For almost a month, my education repositories slowed down.
Not because I had lost interest. The opposite was true. I had built so much that I could keep adding features without knowing whether I was improving the experience.
AlgoQuest had Hero Books. Algorithm Builder could forge artifacts. Qbit could accompany a mission. Scholarium was becoming the meeting place for the suite.
But a child would not arrive with five years of SecuredMe context already in mind. They would open a mission, enter a workshop, ask an agent for help, close the laptop, and return later.
At that exact moment, who remembers what they chose? Which tool owns progression? How does the child distinguish a suggestion from a decision?
Git records commits. It does not record the moment an architect looks at his own system and admits: I may be too close to see what is missing.
So I stopped building inside the suite long enough to learn somewhere else.
WebMCP made the button honest
The WebMCP hackathon became my first laboratory.
I went there thinking about side panels. I came back thinking about authority.
A side panel is the visible room beside the work. WebMCP asks the harder question: what named tool exists behind the button, what it accepts, what it returns, and where its power ends.
For a human, a button labelled "Continue" may feel obvious. For an agent, it is incomplete.
Continue which mission? From which revision? Is the action reading, staging, or executing? Who approves it?
That changed my design vocabulary. I stopped asking only, "Can the companion call the function?" and started asking, "Can the system explain the authority, evidence, change, and return path?"
WebMCP is still experimental, but the lesson is already practical: an agentic interface begins with clear promises around each action.
Thanks2Go made the boundary real
The DEV Weekend Challenge became my second laboratory.
With Thanks2Go, I wanted something small: a voluntary two-dollar gesture of gratitude. Not a payment platform. Just a way to say: your help mattered.
The mechanism had to remain smaller than the gesture.
Intent is not approval. Approval is not provider acceptance. A cryptographic signature is not confirmed settlement.
So Thanks2Go used visible states. The agent could inspect and prepare. The irreversible crossing remained a visible human act.
Then another consenting person used the live PayPal path. Two dollars were sent. The transaction completed. I received 1.64 USD after fees.
That was small. It was also complete.
It did not validate every failure path. It did not turn Thanks2Go into a financial product. It proved exactly one live success path.
That discipline returned with me to the Hero Books.
A specialist suggestion is not the child's decision. A returned artifact is not admitted evidence. A replayed request must not advance the same mission twice.
The companion is not a tool menu
Before this pause, I thought of the companion mainly as quick access to tools.
That definition no longer works.
For a Hero Book, the panel must become a living logbook: hero, world, mission, step, talents, inventory, decisions, dice, evidence, tokens, and active specialist.
This is not decoration. It is where a child recovers the continuity of an adventure.
That continuity cannot depend on a panel staying open. Manifest V3 service workers are ephemeral. A memory variable can vanish. A laptop can close at the wrong moment.
So I defined HeroBookPanelState.v1 as explicit, versioned, and revision-aware. The cockpit can reopen the experience, but it does not own progression. AlgoQuest remains canonical.
Algorithm Builder is the forge. Qbit carries continuity. Codex or Gemini can help as specialists.
AlgoQuest keeps the story.
The specialist returns a proposal or receipt. AlgoQuest decides what enters the mission.
The pause became the architecture
I also learned not to outsource proof to a grade.
webmcp.com's evaluation helped me. It found weaknesses and gave direction. But a third-party score measures its own methodology.
So I built a local Evidence Gate in Scholarium. It checks names, schemas, modes, handlers, permissions, refusals, receipts, and agent journeys.
That proves what its own code can inspect.
Only that.
To know whether the cockpit helps a child find their way, I still need real evaluation with children, parents, educators, and accessibility users.
The companions are not finished. Builder still needs to feel like a forge. AlgoQuest needs the rhythm of a gamebook. Qbit needs presence.
But the work is no longer vague.
The engine can read without changing, stage without claiming execution, wait for approval, reject a stale revision, issue a receipt, and return authority.
These behaviors are less spectacular than an avatar or a 3D scene. They are what future surfaces can rest on.
We spend enormous effort teaching AI when to stop. Before payment, publication, or canonical mutation, we ask for confirmation.
The human architect needs the same discipline.
Nobody returns HUMAN_ARCHITECT_APPROVAL_REQUIRED for me.
I must recognize the signal.
I do not want the largest possible number of agentic tools. I want a child moving from AlgoQuest to Builder and then to a specialist to recover their hero, mission, decisions, creations, and earned evidence.
Eventually, I want the companion to help the child need the companion less.
Knowing how to move forward is a skill.
Knowing where to stop is another.
Explore
- WebMCP proposal: https://github.com/webmachinelearning/webmcp
- Chrome WebMCP docs: https://developer.chrome.com/docs/ai/webmcp
- PayPal idempotency: https://developer.paypal.com/reference/guidelines/idempotency/
Top comments (0)